| # Part-specific specialists | |
| Confirmed initial release: six historical JC specialists and six historical SK specialists. Each bank has one model per target part. These historical models do not require enrollment audio. Checkpoint paths and SHA-256 hashes are listed in [manifest.json](../manifest.json); weight license: CC BY-NC 4.0. | |
| The later singer-enrollment specialist experiments are excluded from the public | |
| release. The [public inference code](https://github.com/pting126/ACappellaSet/blob/main/docs/inference.md#original-fixed-and-part-specific-models) supports these original part-specific banks with verified local checkpoints. Short-input CPU parity against the native implementation has been verified; the inference-only exports preserve every model tensor and omit optimizer state. | |